Defect Dynamics in 2‑D MoS

Probed by Using

Machine Learning, Atomistic Simulations, and

High-Resolution Microscopy

Tarak K. Patra,

†

Fu Zhang,

‡,§

Daniel S. Schulman,

‡

Henry Chan,

†

Mathew J. Cherukara,

†,⊥

Mauricio Terrones,

‡,§,#,⊗

Saptarshi Das,*

,§,∇

Badri Narayanan,*

,∥,¶

and Subramanian K. R. S. Sankaranarayanan*

,†,&

†

Center for Nanoscale Materials,

∥

Materials Science Division, and

⊥

X-ray Science Division, Argonne National Laboratory, Argonne,

Illinois 60439, United States

‡

Materials Science and Engineering,

Center for 2-Dimensional and Layered Materials,

Department of Physics,

⊗

Department of

Chemistry, and

∇

Department of Engineering Science and Mechanics, The Pennsylvania State University, University Park,

Pennsylvania 16802, United States

Computational Institute, University of Chicago, Chicago, Illinois 60637, United States

Supporting Information

ABSTRACT: Structural defects govern various physical, chemical, and

optoelectronic properties of two-dimensional transition-metal dichalcoge-

nides (TMDs). A fundamental understanding of the spatial distribution and

dynamics of defects in these low-dimensional systems is critical for advances

in nanotechnology. However, such understanding has remained elusive

primarily due to the inaccessibility of (a) necessary time scales via standard

atom istic simulations and (b) required spatiotemporal resolution in

experiments. Here, we take advantage of supervised machine learning, in

situ high-resolution transmission electron microscopy (HRTEM) and

molecular dynamics (MD) simulations to overcome these limitations. We

combine genetic algorithms (GA) with MD to investigate the extended

structure of point defects, their dynamical evolution, and their role in

inducing the phase transition between the semiconducting (2H) and

metallic (1T) phase in monolayer MoS

. GA-based structural optimization

is used to identify the long-range structure of randomly distributed point defects (sulfur vacancies) for various defect

densities. Regardless of the density, we ﬁnd that organization of sulfur vacancies into extended lines is the most

energetically favorable. HRTEM validates these ﬁndings and suggests a phase transformation from the 2H-to-1T phase

that is localized near these extended defects when exposed to high electron beam doses. MD simulations elucidate the

molecular mechanism driving the onset of the 2H to 1T transformation and indicate that ﬁnite amounts of 1T phase can

be retained by increasing the defect concentration and temperature. This work signiﬁcantly advances the current

understanding of defect structure/evolution and structural transitions in 2D TMDs, which is crucial for designing

nanoscale devices with desired functionality.

KEYWORDS: machine learning, microscopy, atomistic simulations, 2D materials, phase transitions, and defect dynamics

ransition-metal dichalcogenides (TMDs) exhibit ver-

satile electronic properties and, in turn, have attracted

tremendous attention for possible applications in

nanoscale electronic devices.

1−4

In these 2D-TMDs, a myriad

of structural defects can either pre-exist or be introduced

during sample preparation, processing, and transfer processes.

These defects can be points (e.g., chalcogen vacancies) or of

extended kind ( e.g., dislocations and boundaries); they are

known to have a profound inﬂuence on the physicochemical,

electronic, and optical properties of these materials.

5,6

For

example, electrical-transport measurements on thin sheets of

MoS

almost universally reveal n-type ﬁeld eﬀect transistor

(FET) characteristics, largely owing to the presence of

chalcogen vacancies, impurities, and metal-like antisite defects

that pin the Fermi level of the metal at the metal/TMD

contact interfaces.

7−10

In addition, a high defect density is

strongly correlated with low carrier mobility

11−13

as well as

Received: April 16, 2018

Accepted: August 3, 2018

Published: August 3, 2018

Article

www.acsnano.org

Cite This: ACS Nano 2018, 12, 8006−8016

ACS Nano 2018, 12, 8006−8016

Downloaded via BINGHAMTON UNIV STATE UNIV NEW YORK on April 20, 2023 at 18:05:10 (UTC).

See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

hysteresis in FET characteristics, since they act as adsorption

sites for chemical spe cies. While defects are considered

detrimental for conventional high-performance electronic

devices, precise control over the density and distribution of

defects may open up avenues for electronic applications, such

as neuromorphic devices.

14,15

Previous experimental and theoretical studies suggest that

the most abundant defect in TMDs, such as MoS

, is the

chalcogen (sulfur) monovacancies. These point defects often

reorganize to form extended structures during sample

processing, transfer, as well as operation of a device; atomic-

scale dynamics of these extended defect structures control key

structural transitions.

Of particular interest is the defect-

mediated crossover between 2H (semiconductor) and 1T

(metallic) phases of MoS

mainly owing to its immediate

relevance to nanoscale device applications.

17−19

The reorgan-

ization of point defects necessary for 2H-to-1T transition can

be triggered via operating/processing conditions of heating,

prolonged-electron beam irradiation, or lithiation.

17−19

Struc-

turally, 2H phase possesses a trigonal prismatic arrangement of

a molybdenum (Mo) atom sandwiched between two sulfur (S)

atoms, while the 1T phase exhibits an octahedral coordina-

tion.

In addition to diﬀerent electronic properties, the two

phases also exhibit contrasting magnetic behavior; 1T phase is

paramagnet, while 2H phase is weakly diamagnetic.

A precise

control over this phase transition is of great interest to make

speciﬁc spatial patterns of 1T/2H domains, such as placing low

resistance 1T phase electrical contacts with an atomically sharp

interface to the semiconducting 2H channel in FETs.

The

2H and 1T phase can coexist in a single-layer MoS

, and the

relative fractions can be varied to tune their optoelectronic and

electrocatalytic properties. For example, when the contents of

the 1T phase increase, the photoelectric conversion eﬃciency

of a MoS

-based photo anode increases,

the performance of a

MoS

based supercapacitor improves,

and the eﬃciency of

nanosheets for hydrogen evolution reactions (HER) is

enhanced.

Interestingly, the defect distribution in a Li-

intercalated MoS

sheet can also be altered by sonication

and

solvent thermal treatment to reverse the transformation from

the 1T to the 2H phase.

Evidently, a fundamental

understanding of the atomic-scale structure and dynamics of

defects in 2D TMDs as well as their role in phase transitions is

very crucial for device applications. However, such an

understanding has remained elusive until now.

The transition between 2H and 1T phases involves local

rearrangement of S atoms with respect to their central Mo

atom. A previous study on Re-doped MoS

reports that the

2H-to-1T transition proceeds via formation (and subsequent

intersection) of an intermediate alpha (α) phase under

electron-beam radiation at high temperatures.

Although

this work characterized the structure of α-phase to contain a

shrunk zig-zag pattern of Mo atoms,

the key question as to

why such an intermediate phase is required for the 2H to 1T

phase transformation remains unanswered. Recent density

functional theory works have shown that the 2H-to-1T phase

transition is associated with a ﬁnite energy barrier, with 2H

being the energetically most favorable phase;

they postulate

that the electron transfer between MoS

and the dopants

initiate the 2H-to-1T transformation.

However, the molec-

ular mechanisms underlying the 2H-to-1T transition, the role

of defects, as well as the dynamical evolution/rearrangement of

S and Mo atoms during this phase transformation have not yet

been established. The major challenge in addressing these

questions is that the atomic-scale processes governing

formation of extended defects, and phase transition involves

a broad range of time scales from picoseconds to several

seconds or minutes. The shorter time scales (picoseconds) can

be accessed via ultrafast synchrotron X-ray experiments, such

as coherent diﬀraction imaging. Although our recent study

using coherent diﬀraction imaging at state-of-the-art synchro-

tron facilities has been successful in investigating multilayered

2-D dichalcogenides,

the spatial resolution is limited to 10

Figure 1. Evolutionary search for optimal extended defect conﬁgurations at various concentrations of S vacancies, achieved using genetic

algorithms. (a) Potential energy of the most stable conﬁguration is plotted as a function of generation at various defect concentrations. The

defect concentration is evaluated as the fraction of the available S sites that are vacant. The initial and ﬁnal conﬁgurations of four evolutions

are shown in (b), (c), (d), and (e) for concentrations of 1.5%, 4.0%, 5.0%, and 7.5%, respectively.

ACS Nano Article

DOI: 10.1021/acsnano.8b02844

ACS Nano 2018, 12, 8006−8016

8007

nm, which makes it diﬃcult to track point defects. High-

resolution transmission electron microscopy (HRTEM), on

the other hand, possesses the necessary resolution to image

point defects; but cannot capture ultrafast dynamics owing to

its slow frame speed (∼10 Hz). Standard atomistic simulations

(e.g., molecular dynamics) have atomic resolution but are

limited to several tens of nanoseconds.

Here, we use machine learning, molecular simulations, and

high-resolution transmission electron microscopy (HRTEM)

to overcome these limitations. We investigate the energetics of

arrangement of point defects into extended line defects at

various S-vacancy concentrations and subsequently elucidate

the role of these extended defects in the 2H-to-1T phase

transformation. A genetic algorithm (GA) is used to eﬃciently

search the most energetically favorable distribution of atomic

defects, wherein the energies of individual candidate structures

(genes) are calculated using a reactive force ﬁeld. The stable

structures identiﬁed by our evolutionary search are exper-

imentally veriﬁed via HRTEM. We then investigate the gliding

and rearrangement of S and Mo atoms around the regions of

extended defects in the MoS

layer via molecular dynamics

(MD) simulations. These MD simulations identify the

molecular mechanisms governing α-phase formation near the

exten ded defects in 2H-MoS

and its inﬂuence on the

subsequent transition to the 1T phase. Furthermore, our

large-scale MD simulations demonstrate the impact of

temperature on the size, shape, and concentration of 1T

phase domains in the monolayer. In addition, we observe that

such a transformation is absent in randomly distributed S-

vacancies in a MoS

layer. Overall, our study elucidates atomic-

scale dynamics associated with defect migration, provides a

mechanism for 2H-to-1T transformation, and outlines useful

design rules for engineering the properties of 2D materials by

deviating from their stoichiometric composition.

RESULTS AND DISCUSSION

Evolutionary Structural Optimization of Defects in

2D MoS

. We ﬁrst track the extended conﬁgurations of sulfur

vacancies in a monolayer MoS

by performing GA-based

structural optimization. Our search for extended conﬁgurations

made up entirely of S vacancies is well-founded in light of

recent experiments, which indicate that S vacancies are the

most dominant type of defects in MoS

sheets obtained via

mechanical exfoliation and chemical vapor deposition.

Furthermore, previous DFT calculations report that the

formation energies of S-vacancy (2.12 eV) are lower than all

other types of defects including anti-sites.

As mentioned here,

we start our GA runs with initial populations of candidates

containing randomly distributed S-vacancies at the desired

vacancy density. Figure 1 shows the evolution of defect

structures from randomly dispersed vacancies to ordered

extended conﬁgurations in a MoS

monolayer at four diﬀerent

S-vacancy concentrations (ρ = 1.5, 4, 5 and 7.5%). The

potential energy of the most stable structure obtained at each

generation is plotted in Figure 1a, which converges in ∼600

generations for all of the cases. Parts b−eofFigure 1 depict the

atomic snapshots of initial randomly distributed S-vacancies

and ﬁnal optimized conﬁgurationsatvariousvacancy

concentrations. During the early stages of evolution (within

ﬁrst ∼100 generations), we observe formation of sm all

aggregates of vacancies, such as dimers and trimers. The

energetic gain in forming these aggregates is ∼0.001 eV/atom

for ρ = 1.5% and 0.018 eV/atom for the highest concentration

studied, i.e., ρ = 7.5%. The number of dimers and trimers is a

Figure 2. High-resolution TEM images showing the distribution of line defects in MoS

samples with diﬀerent defect densities: (a) light, (b)

medium and (c) heavy. Extended line defects identiﬁed by our GA search are seen in this HRTEM images of MoS

. Defect distribution under

electron beam (∼50000 e

−

/Å

·s) are shown in (d) and (e) at time t = 0 s and t = 120 s, respectively. (f) Atomic-resolved HAADF-STEM

image of monolayer MoS

shows various defect conﬁgurations.

ACS Nano Article

DOI: 10.1021/acsnano.8b02844

ACS Nano 2018, 12, 8006−8016

8008

strong function of the defect density and shows a non-

monotonic trend; the highest fraction of vacancies occurs in

the form of dimers/trimers (∼3%) for ρ = 7.5% after ∼130

generations. As the GA run proceeds further, we observe that

most of the vacancies exist as small line defects (∼3−4S

vacancies). Depending on the defect densities, we also observe

a rich variety of conﬁgurations, including triangular shapes

(Figure 1d) or zig-zag (Figure 1e). This is consistent with

previous experiments, which report the existence of triangular

holes,

32,33

ﬂower shapes,

and amorphous networks,

along

with lines in electron-beam irradiated 2D materials. Addition-

ally, triangular holes formed by three missing neighboring

sulfur atoms (Figure 1d) suggest the possibility of a V

MoS3

phase in a MoS

layer as seen in our HRTEM; both of these

ﬁndings are consistent with previous experimental works.

The

lowest energy structures obtained at the end of the GA runs

reveal extended line defects of sulfur vacancies (>10) to be the

most favorable at all vacancy concentrations considered in this

work (Figure 1b−e). The energies of the optimized defect

structures are 5.0034, 5.01844, 5.0195, and 5.00866 eV/atom

lower than the initial random population for ρ = 1.5%, 4.0%,

5.0%, and 7.5%, respectively.

The energy of the most stable structure is expected to

increase monotonically with defect concentration. Indeed, the

lowest energy of GA-optimized conﬁgurations, for ρ ≥ 4%,

follow this intuitive trend. At the lowest defect concentration

studied here (i.e., ρ = 1.5%,), however, the converged energy

value is higher than all of th e oth er cases, which is

counterintuitive. The GA-identiﬁed lowest energy conﬁgu-

ration (after 600 generations) for ρ = 1.5% consists of a

combination of lines (of varying sizes) and isolated vacancies

(cf. Figure 1b), which results in a somewhat higher energy

∼−5.0034 eV/atom. Ordering all of the vacancies in a single

line, for ρ = 1.5%, results in a drop of ∼17 meV/atom, yielding

an energy value of −5.021 eV/atom, lower than that of higher

defect concentrations (ρ ≥ 4%). Thus, the GA run for ρ =

1.5% does not reach the global energy minimum within 600

generations. Nevertheless, it is important to note that the

energetic diﬀerence between the conﬁguration shown in Figure

1b and the global minimum (∼ 17 meV/atom) is within the

thermal ﬂuctuations at room temperature. Thus, extended

defect conﬁgurations containing lines of diﬀerent lengths (such

as that identiﬁed by GA in Figure 1b) are all likely to occur at

room temperature, alongside the global minimum structure.

This ﬁnding is ﬁrmly supported by our HRTEM images

(Figure 2), which show a range of defect conﬁgurations with

lines of varying lengths, along with isolated defects. We further

note that the slow convergence of GA for low defect

concentration case (i.e., extremal cases) is typical of most

ML methods.

High-ResolutionElectronMicroscopyonDefect

Evolution and Distribution in MoS

Samples. We

Figure 3. Role played by line defects in 2H-to-1T phase transition in MoS

. (a) Energy barrier associated with the 2H-1T transition.

(Inset)

Top and side view of the two phases are shown schematically. (b) Initial conﬁguration of a line defect in a 2H-MoS

layer of dimension 5.8

nm × 5.2 nm. The line defect is shaded for clarity. (c) Evolution of Mo atoms proximal to line defect at 900 K. (d) S atoms in one plane glide

with respect to other leading to rapid shrinking of Mo−Mo bonds near the line defect forming an intermediate α-phase. (e) Side view of S

atoms gliding over longer time scales results in 1T phase. Mo atoms are shown in red whereas atoms in yellow represent top layer of S and

atoms in blue represent the bottom layer of S.

ACS Nano Article

DOI: 10.1021/acsnano.8b02844

ACS Nano 2018, 12, 8006−8016

8009

performed detailed characterization of the defect distribution

in 2D MoS

to validate the structural predictions made by our

GA. Three electro ablated monolayer MoS

region s with

diﬀerent defect densities were imaged by HRTEM with a

large overfocus imposed to ensure the bright atom contrast and

easily distinguish the defects from perfect hexagonal lattice.

The structural defects evolve from randomly distributed sulfur

monovacancies (light defect density shown in Figure 2a) and

distributed line defects (medium density shown in Figure 2b)

to extended coupled line defect induced structural disorders

(heavily defective structure shown in Figure 2c). Suc h

HRTEM observations testify the GA prediction at diﬀerent

defect densities in 2D MoS

and indicate a large variation in

the size of extended defects. We note that these defects are

intrinsic in the sample as the doses used for beam shower and

imaging are mild (∼5000 e

−

/Å

·s) and unlikely to create

knock-on damages. However, prolonged electron irradiation in

the monochromated TEM mode (∼50000 e

−

/Å

·s for 120s,

which is 1-fold higher than the imaging dose, demonstrated in

Figure 2a−c) generates sulfur monovacancies due to the

knock-on damage.

This prolonged electron irradiation helps

randomly distributed sulfur monovacancies to evolve into

extended s tructural imperfections such as line defects,

triangular holes, and amorphous networks as shown in Figure

2d,e. Further, an atomic-resolution high-angle annular dark

ﬁeld scanning TEM (HAADF-STEM) image of the irradiated

monolayer MoS

in Figure 2fshowsvariousdefect

conﬁgurations such as sulfur monovacancies V

and triangular

defects V

MoS3

. Figure 2f also suggests the formation of an α-

phase due to extensive line defects, and coupling of two α-

phase regions at 60° to each other may trigger the formation of

1T domains in the 2H phase of MoS

Atomistic Simulations of Defect-Dri ven 2H-to-1T

Phase Transformation. Our HRTEM characterization

reveals a rich distribution of extended defects and hints at a

possible correlation between the defect densities and the

observed phases (2H, 1T and α). This correlation is also

observed for many other 2D materials.

Here, we aim to

understand this correlation for MoS

monolayers. The 2H

phase of MoS

exhibits ABA staking of atoms, while the atoms

are in an ABC staking in the 1T phase as shown in Figure 3a.

Both DFT a nd ReaxFF calculations suggest an energy

diﬀerence of 0.5 eV per formula unit that exists between the

two phases, with 1T and 2H being the metastable and ground

states, respectively. Further, Nudged Elastic Band calculations

shown in Figure 3a indicate that an energy barrier of 1.25 eV

per formula unit needs to be surmounted in order to transform

from a 2H phase to a 1T phase.

On the other hand, the

Figure 4. Phase transformation on a MoS

layer with randomly distributed line defects. (a) Atomic snapshot after 1 ns of MD simulation at

900 K. The top view of the sheet is shown where blue regions represent extended defects, and yellow beads are the S atoms in the top layer.

Two zoomed regions in (a) are highlighted in (b) and (c) with atomic details. The coexistence of 2H, 1T, and α phases are demonstrated in

(b) where solid lines are guide to the eye. Further, 1T phases are shaded for clarity in (b) and (c). (d) Probability distribution of bonds

between Mo atoms and top layer S atoms at 900 K, where the bond lengths corresponding to 1T, 2H, and α are shown by arrow marks. (e)

Fraction of 1T, 2H, and α phases as a function of temperature.

ACS Nano Article

DOI: 10.1021/acsnano.8b02844

ACS Nano 2018, 12, 8006−8016

8010

energy barrier for the 1T →2H phase transformation is 0.75 eV

per formula unit. Such barriers can be surmounted via heat

treatment, which in turn can yield coexistence of 2H and 1T

phases in the same sample.

19,38,39

This is especially important

considering the widely diﬀerent electronic properties of the 2H

and 1T phases. Previous studies suggest that the trans-

formation between the 2H and 1T phase requires gliding of S

atoms in one plane with respect to the other; such gliding can

be achieved by thermal activation, and exposure to an electric

beam.

Here, we explore the eﬀects of extended line defects

seen in our GA and HRTEM on the phase transformation

between the 2H and 1T phases. In particular, we perform MD

simulations to better understand how the gliding of atoms

takes place during the transformation and gain insights into the

atomic scale pathways via which the energy barrier is

surmounted in a MoS

layer during heat treatment. We ﬁrst

perform a series of controlled MD simulations on pristine

phases. Our MD simulations on the pristine 1T phase and 2H

phase do not show any transformation, regardless of the

temperature (cf. Figure S1, Supporting Information). Fur-

thermore, randomly distributed atomic defects on these two

phases do not induce a phase transformation (cf. Figure S2,

Supporting Information). Next, we perform MD simulations

on conﬁgurations with extended line defects. These simu-

lations demonstrate that phase transformation initiates only

when extended defects appear on a MoS

layer, which is

identiﬁed as the lowest energy conﬁguration by our GA study.

We performed a series of MD simulations at diﬀerent

temperatures starting with conﬁgurations that have randomly

distributed line defect(s) (similar to GA predictions) in a

MoS

layer. Parts b−eofFigure 3 illustrate the atomic-scale

mechanisms for the onset of the 2H-to-1T transition in a MoS

sheet with a single line defect, as identiﬁed by our MD

simulations. The initial conﬁguration is entirely 2H containing

a line defect (Figure 3b), with a conﬁguration similar to that

obtained from our evolutionary structural searches (Figure

1b−e). The temporal evolution of the region proximal to the

line defect is shown in Figure 3 c. We observe two types of

phase transformations that have widely diﬀerent time scales.

There is an initial rapid formation of the so-called α phase near

the line defect within the ﬁrst 10 ps of the MD run. During this

α-phase formation, top S atoms from both sides of the line

defect glide toward the defect, as shown by the solid black

arrows in Figure. 3. The gliding of S atoms is concurrent with

the shrinking of Mo atomic lines in the middle layer of the

sheet (Figure 3d). Over much longer time scales (∼1 ns), the

1T phase starts to nucleate near the α region, which is

highlighted by shaded regions in the bottom panel of Figure

3c. In addition, the gliding of S atoms during the 1T phase

formation (side view) is shown by solid arrows in Figure 3e.

Such coexistence of the 2H, 1T, and α-phase is also seen in our

HRTEM experiment (cf. Figure 2).

In order to understand the kinetics associated with local

growth of the 1T phase and its stability, we perform large-scale

MD simulations where line defects are randomly placed in a

21.8 nm × 24.9 nm MoS

sheet. This large sheet of MoS

represents a distribution of extended line defects (identiﬁed by

GA) that are in close agreement with HRTEM images shown

in Figure 2. The equilibrium conﬁguration of the large area

MoS

sheet is shown in Figure 4a for a representative

temperature (900 K). The system reaches a steady state within

∼1 ns, wherein 1T, 2H, and α phases coexist (Figure 4a,b).

Similar to the single-line defect system, we also observe

transformation of the 2H phase to 1T phase locally near the

line defects. In particular, discrete linear regions (one-

dimensional) of the 1T phase form next to the α-phase; a

representative region close to a line defect that transforms to

2H-α-1T is shown in Figure 4b. Furthermore, two-dimensional

growth of 1T phase is seen at the intersection of two-line

defects, as shown in Figure 4c. This observation is consistent

with the ﬁndings of our HRTEM images (Figure 2f). Next, we

analyze the distribution and amounts of various phases as a

function of temperature. We use a distance criterion to

distinguish diﬀerent phases after the MoS

layer reaches a

steady state at any given temperature. The Mo and S atoms

belonging to 1T, 2H, and α-phases are distinguished using the

shortest Mo−S separation distances (1T: 2.18 Å, 2H: 2.33 Å

and α: 2.45 Å). The detailed in-equilibrium conﬁgurations and

interatomic distances for various phases in a MoS

monolayer

can be found elsewhere.

20,41,42

The probability distribution of

diﬀerent phases at 900 K is shown in Figure 4d. At 900 K, 14%

of Mo−S bonds correspond to the 2H phase, while 1.7% of the

Mo−S bonds correspond to the 1T phase. A large portion of

Mo−S bonds (26%) remains in the α-phase, which may

transform into the 1T phase over extended time scales not

accessible to MD simulations. Our simulations further suggest

that the relative proportions of the three phases can be

controlled by varying the temperature. For example, the

proportion of the 1T and 2H phase increases with temperature

as shown in Figure 4e. The data suggest that the rate of

increment in 1T phase is greater than that of 2H phase. The

increment in the proportion of the 2H and 1T phase is due to

the reduction of α-phase as temperature increases.

Figure 4e shows that the fraction of 1T phase is low (0.5%−

2.5%) regardless of the temperature; nevertheless, even this

small amount of metal region has a signiﬁcant impact in device

application. It has the potential to signiﬁcantly decrease the

Schottky barrier and strong Fermi level pinning on the

extended defects.

Therefore, creating such a small domain of a

local conducting region is very important for many device

applications such as making a conducting ﬁlament in a large

TMD layer for neuromorphic devices. Furthermore, we expect

that the fraction of 1T phase can be tailored by changing the

distribution of extended defects, and alignment of extended

defects within a particular distribution. Our ﬁndings clearly

demonstrate the possibility of controlling the size and shape of

1T phases by manipulating the defect densities, temperature,

and/or alignment of extended defects.

CONCLUSIONS

Evolution of point defects into extended defects and their role

in inducing phase transformation in an atomically thin sheet of

a TMD layer involves atomic processes that occur over a broad

range of time scales spanning picoseconds to several seconds.

Traditional experimental and simulation methods often fail to

provide a holistic understanding of such multiscale phenom-

ena. Here, we combine machine learning, molecular dynamics

simulations, and high-resolution electron microscopy to fully

understand phase transformation in a TMD layer. In particular,

we investigated the defect distribution and their role in driving

the 2H-to-1T phase transformation using MoS

as a model

TMD material. The genetic algorithm identiﬁes the aggrega-

tion of point defects (0D) into a line defect (1D) on a MoS

layer, which is the energetically most stable structure; such

aggregation of point defects into ordered line defects occurs

regardless of the vacancy concentration. These predictions are

ACS Nano Article

DOI: 10.1021/acsnano.8b02844

ACS Nano 2018, 12, 8006−8016

8011

validated by our in situ HRTEM experiments. The HRTEM

experiments suggest a possible semiconductor (2H) to metal

(1T) phase transformation locally in the vicinity of the line

defect. We conduct MD simulations that elucidate the

molecular mechanism of this d efect-driven phase trans-

formation. Our simulations reveal that the S atoms glide

locally toward the defect sites and lead to the formation of an

intermediate α-phase. Th is α-phase triggers 1T phase

formation locally. We demonstrate that a typical 2H → 1T

phase transformation in a MoS

layer is associated with

phenomena occurring over several diﬀerent time scales: (i) a

long time, typically seconds and minutes, over which atomic

defects migrate and coalescence into extended line defects, (ii)

a rapid transformation of 2H to α-phase (∼10 ps) around

these extended line defects, and (iii) an intermediate time scale

(1 ns), over which the α-phase initiates the 2H → 1T phase

transformation. The coalescence of point defects and

formation of 1T require activation energy, which is typically

provided through heating and electron beam radiation in

experiments. Therefore, by introducing line defects, the relative

proportions and the coexistence of metallic 1T and semi-

conducting 2H phases can be systematically tuned in an

atomically thin MoS

layer. Our HRTEM and MD simulations

suggest that the size and shape of the 1T regions is inﬂuenced

by the alignment of the extended line defects. Overall, the

present study elucidates the defect aggregation and defect

driven phase transition mechanism in 2D TMD materials. This

has potential implications in fabricating electrical circuits in the

semiconducting MoS

monolayer for nanoscale electronic

applications. In addition, our work provides a generic

framework where machine learning can be combined with

molecular simulations and electron microscopy to study

multiscale processes associated with complex materials

phenomena.

METHODS

Over the past few years, machine learning and artiﬁcial intelligence

based methods have advanced various facets of materials science,

including accelerated synthesis via robotics,

enhancing existing

material characterization methods,

searching for thermodynamically

stable structures,

45,46

materials design,

36,47

and developing accurate

atomistic models.

48−50

In the context of 2D materials, such ML-based

techniques have enabled rapid searches for exfoliated 2D monolayers

and their assembly into superlattices/heterostructures

as well as

enhancing optical characterization of large-area 2D materials.

Simulation methods such as ab initio molecular dynamics (AIMD)

simulations, classical MD, and kinetic Monte Carlo simulations can be

used to understand defect dynamics. AIMD simulations due to their

high cost are limited in the time and length scales they can access.

Some of our recent studies push the limit of what can be achieved

with AIMD

51−53

(we note that these simulations were performed on

Argonne’s leadership computing and a single simulation required

10000 cores over 6 months for simulation time scales of few tens of

picoseconds). Clearly, studies of defect dynamics (approximately

several nanoseconds and longer) remain intractable with AIMD.

Kinetic Monte Carlo is a powerful mesoscale technique to access

much larger length scales and longer time scales but relies on

knowledge of a predeﬁned rate catalog; i.e., mechanisms have to be

known a priori.

54,55

Classical MD overcomes these limitations and

allows us to understand extended defect dynamics over nanometer/

nanosecond scales. While there are reports of machine learning

enhanced molecular dynamics (MD) simulations (i.e., the use of

machine learning to develop accurate interatomic potentials for use in

MD simulations), this work uses machine learning to perform

structural optimization and evolution of the point defects into

extended structures. Essentially, this work involves use of GA to

sample conﬁgurations of defect structures to identify ones that are

energetically favorable; in the case of our target system MoS

, the line

defects (ordered vacancies) are more stable than isolated vacancies

(i.e., point defects). For instance, GA has been used to obtain optimal

potential parameter sets for a bond-order potential to describe

interactions in Co−C heterostructures.

Similarly, Csa

nyi et al.

employ machine le arning to c onstruct force-ﬁeld models for

amorphous silicon, while Ramprasad et al.

provide a perspective

on the use of machine learning to develop interatomic potentials for

various materials. In short, these works (and other related report)

involve designing strategies for eﬃcient parameter optimization, while

we are focused on deriving the most stable structural conﬁguration of

extended defects.

It should be noted that structural optimization via GA has also

been used for 2 D science mainly for crystal-structure predic-

tion.

36,47,57−59

In most studies, genetic algorithm (GA) in

combination with density functional theory (DFT) is used to search

the global minimum structures. A typical GA run would involve 20

populations run over 100 generations (total ∼2000 evaluations per

system). Owing to the computational cost of DFT, these studies

involve system sizes of single to few unit cells. In order to use GA to

predict extended defect structures such as the line defects (that

extended to several nanometers), one requires a system size that is

much larger (supercell ∼20 × 18−360 unit cells). This is

computationally intractable with DFT. Empirical force-ﬁelds with

much lower cost allow us to eﬃciently sample the defect

conﬁgurations. Hence, we use GA and MD to understand defect

dynamics of extended structures.

Similarly, th ere have been previous eﬀorts on advancing

microscopy techniques to characterize defect dynamics. For example,

in a seminal work, Hashimoto et al.

show in situ defect formation in

single graphene layers by high-resolution TEM. Meyer et al.

represent a major advance in using TEM to image the dynamics of

light atoms and molecules on graphene. Likewise, Sim et al.

employs

electron microscopy studies to investigate how defect structures in

multilayered 2D materials diﬀer from their monolayers. While such

electron microscopy studies can capture temporal evolution, their

resolution is still limited to frame speeds ∼10 Hz and consequently

cannot capture short-lived events (nanosecond time scales). Such

ultrafa st dynamica l events are cru cial for de fec t- drive n phase

transitions between semiconducting and metallic phases in TMDCs

(e.g., MoS

). Overall, it is not possible for a single technique to probe

the dynamics occurring over disparate time and length scales.

Here, ML methods are combined with molecular dynamics

simulations and model validation is performed via HRTEM

experiments to obtain a holistic picture of how point defects organize

into extended structures, which subsequently drive the 2H-to-1T

transition. This understanding cannot be provided by a single method

alone since the time scales involved in defect dynamics vary from

picoseconds (too fast for HRTEM) to several minutes (too long for

MD). By using HRTEM, GA, and MD simulations, we show the

defect evolution from randomly distributed point defects to extended

line defects and highlight their role in driving the 2H to 1T phase

transitions in 2D MoS

Speciﬁcally, we employ GA to identify

extended defect structures that form via organization of point defects;

the GA-identiﬁed structures are validated by HRTEM. Molecular

dynamics simulations, alongside HRTEM, elucidate (a) that extended

defect structures act as precursors for 2H-to-1T transition and (b) the

coexistence of 2H and 1T phases in a single monolayer. In all machine

learning studies, model validation is an important aspect and HRTEM

serves this important purpose of validating the predictions of GA. In

the subsequent sections, we provide the details of the three

techniques.

Evolutionary Structural Search. Evolutionary searches identify

the global energy minimum structure of a material along with

structures that are energetically close (but higher) than the most

stable structure; often, such structures are kinetically trapped

depending on processing and operating condition of the material.

Therefore, this method is widely used in materials design

problems.

36,62,63

Here, we adopt an evolutionary search, viz., genetic

ACS Nano Article

DOI: 10.1021/acsnano.8b02844

ACS Nano 2018, 12, 8006−8016

8012

algorithms (GA) to identify energetically favorable arrangements of

vacancies in one of the S planes of a MoS

layer. In GA, candidate

materials structures are mapped to a genome upon which the

evolutionary operations are employed. In a single layer MoS

, a plane

of Mo atoms is sandwiched between two planes of S atoms; in each S

plane, the atoms are organized in a 2D triangular lattice. As the

interlayer vacancy migration is associated with a very high energy

barrier (>5 eV), we restrict our search space to the top layer of S

atoms. To describe the vacancy distribution in a given candidate

structure, we deﬁne a 2D binary genome. This genome is a string

containing the current state for each site in the 2D triangular lattice;

each site can be in one of the two states, namely: (a) “0” representing

the absence of a S atom (vacancy) and (b) “1” indicating that site is

occupied by a S atom. For a prescribed vacancy density, the

evolutionary search begins with a random population of 32

candidates, each with an arbitrarily chosen but distinct genome,

(i.e., S-vacancy distributions). The ﬁtness of each candidate is

calculated as the potential energy of the MoS

monolayer containing a

distribution of S-vacancies as deﬁned by its 2D genome. We note that

the candidate structures with lower potential energy possess higher

ﬁtness, and have higher chance of survival in an evolution. The atomic

interactions in the defective MoS

structures are modeled using the

reactive force ﬁeld (ReaxFF)

within LAMMPS

package. In each

generation, genetic operations, namely, selection, mutation, and

crossover, are performed on the current population to generate 32

new candidates; details of these operations are provided else-

where.

36,47

The candidates are then ranked by their ﬁtness, and the

best (ﬁttest) 32 candidates are passed to the next generation. This

procedure is iterated until the GA run converges; i.e., the potential

energy of the ﬁttest candidate does not change over a long period of

time (∼100 generations). We note that several case studies are

conducted with varying number of population size, and 32 is found to

be suﬃciently large to ensure necessary structural diversity in the

population (in the initial stages) as well as convergence to low-energy

conﬁgurations with reasonable computational costs. Our evolutionary

searches using population sizes (>32) have resulted in identical lowest

energy conﬁgurations, as that from runs with 32 candidates in the

gene pool. Moreover, the 2D MoS

sheet used for sampling defect

conﬁgurations is suﬃciently large (5.2 nm × 5.8 nm) to avoid ﬁnite

size eﬀects; using a MoS

sheet with twice the area yielded identical

conﬁgurations after convergence of GA runs.

Molecular Dynamics Simulations. To investigate the atomic-

scale mechanisms underlying early stages of 2H-to-1T phase

transition in monolayer MoS

, we employed isobaric−isothermal

(NPT) MD at various temperatures (300−1500 K) and ambient

pressure. All MD simulations are performed using the LAMMPS

package. Defective monolayer MoS

sheets (up to 21.8 nm × 24.9

nm; 17140 atoms) are prepared by pla cing the GA- derived

energetically favora ble extended vacancy conﬁgurations (most ly

lines) along random orientations in one S-plane. Periodic boundary

conditions are employed in the plane of the MoS

sheet. Similar to the

GA searches, the atomic interactions are described by ReaxFF.

Constant temperature and pressure conditions are maintained using a

Nose

−Hoover thermostat and barostat.

Sample Preparation and High-Resolution Transmission

Electron Microscopy. Single-layer MoS

samples were prepared

through an electrochemical thinning process, known as electro-

ablation (EA), using the procedure reported by Das et al.

In the EA

process, multilayer/bulk MoS

ﬂakes are mechanically exfoliated onto

a conductive TiN substrate where a 1.4 V vs Ag/AgCl potential was

applied for <60 s. During the substrate-assisted, self-limited EA, the

multilayer ﬂakes undergo an electro-oxidation process which ablates

away the bulk layers leaving the bottom monolayer intact. TMDs

produced via EA have shown excellent semiconducting properties in

ﬁeld eﬀect transistor (FET) devices; however, photoluminescence

studies have shown that excessive EA times introduce signiﬁcant

numbers of defects such as sulfur vacancies. Since the ablation

initiates at the edge sites and therefore material ablates ﬁrst from the

perimeter working inward toward the center of the ﬂake, the outer

edges are exposed to the EA process for longer periods of time.

67−69

Hence, an inhomogenous defect density is observed as seen in Figure

2 with the edge regions being more defective than the center. The

electroablated exfoliated MoS

ﬂakes were transferred by an etchant

free transfer method

from the TiN substrate onto a Au Quantifoil

TEM grid for TEM analyses; in this technique, most residues are

washed away by several water baths to avoid contamination.

Aberration corrected scanning/TEM (AC-S/TEM) imaging was

performed on a FEI Titan

60−300 microscope at 80 kV accelerating

voltage with a monochromated gun and spherical aberration corrected

lenses, providing subangstrom resolution. The electron dose of ∼5000

−

/Å

·s was used for HRTEM imaging of monolayer MoS

minimize structural damage and to push away possible contaminants.

The high-angle annular dark ﬁeld (HAADF) detector (Fischione)

which was used for ADF-STEM imaging acquisition had a beam

current of 45pA, beam convergence angle of 30 mrad, and collection

angle of 51−300 mrad. The HRTEM and ADF-STEM images in the

content are further ﬁltered by Gaussian blur function via ImageJ

software.

ASSOCIATED CONTENT

Supporting Information

The Supporting Information is available free of charge on the

ACS Publications website at DOI: 10.1021/acsnano.8b02844.

Details of the MD simulations performed for 2D MoS

(pristine as well as with sulfur defects) (PDF)

AUTHOR INFORMATION

Corresponding Authors

*E-mail: sud70@psu.edu.

*E-mail: badri.narayanan@louisville.edu.

*E-mail: skrssank@anl.gov.

ORCID

Daniel S. Schulman: 0000-0002-0751-0578

Henry Chan: 0000-0002-8198-7737

Mathew J. Cherukara: 0000-0002-1475-6998

Saptarshi Das: 0000-0002-0188-945X

Badri Narayanan: 0000-0001-8147-1047

Subramanian K. R. S. Sankaranarayanan: 0000-0002-9708-

396X

Present Address

B.N.: Department of Mechanical Engineering, University of

Louisville, Louisville KY 40292

Author Contributions

S.D., B.N., and S.K.R.S. conceived this project. T.P. carried out

the GA optimization and MD simulations. F.Z. and D.S.

carried out the experiments. All the authors contributed to the

data analysis. S.K.R.S., B.N., and T.P. wrote this manuscript

with contribution from all the authors. All authors have given

approval to the ﬁnal version of the manuscript.

Notes

The authors declare no competing ﬁnancial interest.

ACKNOWLEDGMENTS

This work was supported by Argonne LDRD-2017-012-N0.

Use of the Center for Nanoscale Materials and the resources of

the Argonne Leadership Computing Facility was also

supported by the U.S. Department of Energy (DOE), Oﬃce

of Science, Oﬃce of Basic Science, under Contract No. DE-

AC02-06CH11357. This research used resources of the

National Energy Research Scientiﬁc Computing Center, a

DOE oﬃce of science user facility supported by the Oﬃce of

Science of the US Department of Energy under Contract No.

DE-AC02-05CH112 31. An award of computer time was

ACS Nano Article

DOI: 10.1021/acsnano.8b02844

ACS Nano 2018, 12, 8006−8016

8013

provided by the Innovative and Novel Computational Impact

on Theory and Experiment (INCITE) program. This research

used resources of the Argonne Leadership Computational

Facility, which is a DOE Oﬃce of Science user facility

supported under Contract No. DE-AC02-06CH11357. T.K.P.

thanks Prof. David Simmons, University of Akron, for

providing the GA code, which is modiﬁed and used for this

work. D.S.S. and S.D. acknowledge support from the Air Force

Oﬃce of Scientiﬁc Research (AFOSR) through the Young

Investigator Program with Grant No. FA9550-17-1-0018. F. Z.

and M. T. acknowledge th e NSF-IUCRC Center for

Atomically Thin Multifunctional Coatings 275 (ATOMIC),

under award #1540018.

REFERENCES

(1) Bhimanapati, G. R.; Lin, Z.; Meunier, V.; Jung, Y.; Cha, J.; Das,

S.; Xiao, D.; Son, Y.; Strano, M. S.; Cooper, V. R.; Liang, L.; Louie, S.

G.; Ringe, E.; Zhou, W.; Kim, S. S.; Naik, R. R.; Sumpter, B. G.;

Terrones, H.; Xia, F.; Wang, Y.; et al. Recent Advances in Two-

Dimensional Mate rials beyond Graphene. ACS Nano 2015, 9,

11509−11539.

(2) Das, S.; Robinson, J. A.; Dubey, M.; Terrones, H.; Terrones, M.

Beyond Graphene: Progress in Novel Two-Dimensional Materials and

van Der Waals Solids. Annu. Rev. Mater. Res. 2015, 45,1−27.

(3) Voiry, D.; Mohite, A.; Chhowalla, M. Phase Engineering of

Transition Metal Dichalcogenides. Chem. Soc. Rev. 2015, 44, 2702−

2712.

(4) Cançado, L. G.; da Silva, M. G.; Ferreira, E. H. M.; Hof, F.;

Kampioti, K.; Huang, K.; Pe

nicaud, A.; Achete, C. A.; Capaz, R. B.;

Jorio, A. Disentangling Contributions of Point and Line Defects in the

Raman Spectra of Graphene-Related Materials. 2D Mater. 2017, 4,

025039.

(5) Zhou, W.; Zou, X.; Najmaei, S.; Liu, Z.; Shi, Y.; Kong, J.; Lou, J.;

Ajayan, P. M.; Yakobson, B. I.; Idrobo, J.-C. Intrinsic Structural

Defects in Monolayer Molybdenum Disulfide. Nano Lett. 2013, 13,

2615−2622.

(6) Addou, R.; Colombo, L.; Wallace, R. M. Surface Defects on

Natural MoS2. ACS Appl. Mater. Interfaces 2015, 7, 11921−11929.

(7) Schulman, D. S.; Arnold, A. J.; Das, S. Contact Engineering for

2D Materials and Devices. Chem. Soc. Rev. 2018, 47, 3037.

(8) McDonnell, S.; Addou, R.; Buie, C.; Wallace, R. M.; Hinkle, C.

L. Defect-Dominated Doping and Contact Resistance in MoS2. ACS

Nano 2014, 8, 2880−2888.

(9) Bampoulis, P.; van Bremen, R.; Yao, Q.; Poelsema, B.; Zandvliet,

H. J. W.; Sotthewes, K. Defect Dominated Charge Transport and

Fermi Level Pinning in MoS2/Metal Contacts. ACS Appl. Mater.

Interfaces 2017, 9, 19278−19286.

(10) Das, S.; Chen, H.-Y.; Penumatcha, A. V.; Appenzeller, J. High

Performance Multilayer MoS2 Transistors with Scandium Contacts.

Nano Lett. 2013, 13, 100−105.

(11) Yu, Z.; Pan, Y.; Shen, Y.; Wang, Z.; Ong, Z.-Y.; Xu, T.; Xin, R.;

Pan, L.; Wang, B.; Sun, L.; Wang, J.; Zhang, G.; Zhang, Y. W.; Shi, Y.;

Wang, X. Towards Intrinsic Charge Transport in Monolayer

Molybdenum Disulfide by Defect and Interface Engineering. Nat.

Commun. 2014, 5, 5290.

(12) Qiu, H.; Xu, T.; Wang, Z.; Ren, W.; Nan, H.; Ni, Z.; Chen, Q.;

Yuan, S.; Miao, F.; Song, F.; Long, G.; Shi, Y.; Sun, L.; Wang, J.;

Wang, X. Hopping Transport through Defect-Induced Localized

States in Molybdenum Disulphide. Nat. Commun. 2013, 4, 2642.

(13) Zhu, W.; Low, T.; Lee, Y.-H.; Wang, H.; Farmer, D. B.; Kong,

J.; Xia, F.; Avouris, P. Electronic Transport and Device Prospects of

Monolayer Molybdenum Disulphide Grown by Chemical Vapour

Deposition. Nat. Commun. 2014, 5, 3087.

(14) Arnold, A. J.; Razavieh, A.; Nasr, J. R.; Schulman, D. S.;

Eichfeld, C. M.; Das, S. Mimicking Neurotransmitter Release in

Chemical Synapses via Hysteresis Engineering in MoS2 Transistors.

ACS Nano 2017, 11, 3110−3118.

(15) Sangwan, V. K.; Lee, H.-S.; Bergeron, H.; Balla, I.; Beck, M. E.;

Chen, K.-S.; Hersam, M. C. Multi-Terminal Memtransistors from

Polycrystalline Monolayer Molybdenum Disulfide. Nature 2018, 554,

500−504.

(16) Lin, Y.-C.; Dumcenco, D. O.; Huang, Y.-S.; Suenaga, K. Atomic

Mechanism of the Semiconducting-to-Metallic Phase Transition in

Single-Layered MoS

. Nat. Nanotechnol. 2014, 9, 391.

(17) Friedman, A. L.; Hanbicki, A. T.; Perkins, F. K.; Jernigan, G. G.;

Culbertson, J. C.; Campbell, P. M. Evidence for Chemical Vapor

Induced 2H to 1T Phase Transition in MoX 2 (X = Se, S) Transition

Metal Dichalcogenide Films. Sci. Rep. 2017, 7, 3836.

(18) Han, S. W.; Park, Y.; Hwang, Y. H.; Jekal, S.; Kang, M.; Lee, W.

G.; Yang, W.; Lee, G.-D.; Hong, S. C. Electron Beam-Formed

Ferromagnetic Defects on MoS

Surface along 1 T Phase Transition.

Sci. Rep. 2016, 6, 38730.

(19) Enyashin, A. N.; Yadgarov, L.; Houben, L.; Popov, I.;

Weidenbach, M.; Tenne, R.; Bar-Sadan, M.; Seifert, G. New Route

for Stabilization of 1T-WS2 and MoS2 Phases. J. Phys. Chem. C 2011,

115, 24586−24591.

(20) Ostadhossein, A.; Rahnamoun, A.; Wang, Y.; Zhao, P.; Zhang,

S.; Crespi, V. H.; van Duin, A. C. T. ReaxFF Reactive Force-Field

Study of Molybdenum Disulfide (MoS2). J. Phys. Chem. Lett. 2017, 8,

631−640.

(21) Yan, S.; Qiao, W.; He, X.; Guo, X.; Xi, L.; Zhong, W.; Du, Y.

Enhancement of Magnetism by Structural Phase Transition in MoS2.

Appl. Phys. Lett. 2015, 106, 012408.

(22) Kappera, R.; Voiry, D.; Yalcin, S. E.; Branch, B.; Gupta, G.;

Mohite, A. D.; Chhowalla, M. Phase-Engineered Low-Resistance

Contacts for Ultrathin MoS2 Transistors. Nat. Mater. 2014, 13,

1128−1134.

(23) Pi, Y.; Li, Z.; Xu, D.; Liu, J.; Li, Y.; Zhang, F.; Zhang, G.; Peng,

W.; Fan, X. 1T-Phase MoS2 Nanosheets on TiO2 Nanorod Arrays:

3D Photoanode with Extraordinary Catalytic Performance. ACS

Sustainable Chem. Eng. 2017, 5, 5175−5182.

(24) Jiang, L.; Zhang, S.; Kulinich, S. A.; Song, X.; Zhu, J.; Wang, X.;

Zeng, H. Optimizing Hybridization of 1T and 2H Phases in

MoS2Monolayers to Improve Capacitances of Supercapacitors.

Mater. Res. Lett. 2015, 3, 177 −183.

(25) Voiry, D.; Yamaguchi, H.; Li, J.; Silva, R.; Alves, D. C. B.;

Fujita, T.; Chen, M.; Asefa, T.; Shenoy, V. B.; Eda, G.; Chhowalla, M.

Enhanced Catalytic Activity in Strained Chemically Exfoliated WS

Nanosheets for Hydrogen Evolution. Nat. Mater. 2013, 12, 850−855.

(26) Fan, X.; Xu, P.; Zhou, D.; Sun, Y.; Li, Y. C.; Nguyen, M. A. T.;

Terrones, M.; Mallouk, T. E. Fast and Efficient Preparation of

Exfoliated 2H MoS2 Nanosheets by Sonication-Assisted Lithium

Intercalation and Infrared Laser-Induced 1T to 2H Phase Reversion.

Nano Lett. 2015, 15, 5956−5960.

(27) Sim, D. M.; Han, H. J.; Yim, S.; Choi, M.-J.; Jeon, J.; Jung, Y. S.

Long-Term Stable 2H-MoS2 Dispersion: Critical Role of Solvent for

Simultaneous Phase Restoration and Surface Functionalization of

Liquid-Exfoliated MoS2. ACS Omega 2017, 2, 4678−4687.

(28) Fair, K. M.; Ford, M. J. Phase Transitions and Optical

Properties of the Semiconducting and Metallic Phases of Single-Layer

MoS 2. Nanotechnology 2015, 26, 435705.

(29) Sun, X.; Wang, Z.; Li, Z.; Fu, Y. Q. Origin of Structural

Transformation in Mono- and Bi-Layered Molybdenum Disulfide. Sci.

Rep. 2016, 6, srep26666.

(30) Cherukara, M. J.; Schulmann, D. S.; Sasikumar, K.; Arnold, A.

J.;Chan,H.;Sadasivam,S.;Cha,W.;Maser,J.;Das,S.;

Sankaranarayanan, S. K. R. S.; Harder, R. J. Three-Dimensional

Integrated X-Ray Diffraction Imaging of a Native Strain in Multi-

Layered WSe2. Nano Lett. 2018, 18, 1993−2000.

(31) Komsa, H.-P.; Krasheninnikov, A. V. From Point to Line

Defects in Two-Dimensional Transition Metal Dichalcogenides:

Insights from Transmission Electron Microscopy and First-Principles

Calculations. In GraphITA: Carbon Nanostructures; Springer: Cham,

2017; pp 71−85.

ACS Nano Article

DOI: 10.1021/acsnano.8b02844

ACS Nano 2018, 12, 8006−8016

8014

(32) Jin, C.; Lin, F.; Suenaga, K.; Iijima, S. Fabrication of a

Freestanding Boron Nitride Single Layer and Its Defect Assignments.

Phys. Rev. Lett. 2009, 102, 195505.

(33) Meyer, J. C.; Chuvilin, A.; Algara-Siller, G.; Biskupek, J.; Kaiser,

U. Selective Sputtering and Atomic Resolution Imaging of Atomically

Thin Boron Nitride Membranes. Nano Lett. 2009, 9, 2683−2689.

(34) Kurasch, S.; Kotakoski, J.; Lehtinen, O.; Ska

kalova

, V.; Smet, J.;

Krill, C. E.; Kras heninnikov, A. V.; Kaiser, U. Atom-by-Atom

Observation of Grain Boundary Migration in Graphene. Nano Lett.

2012, 12, 3168−3173.

(35) Kotakoski, J.; Krasheninnikov, A. V.; Kaiser, U.; Meyer, J. C.

From Point Defects in Graphene to Two-Dimensional Amorphous

Carbon. Phys. Rev. Lett. 2011, 106, 105505.

(36) Patra, T. K.; Meenakshisundaram, V.; Hung, J.-H.; Simmons,

D. S. Neural-Network-Biased Genetic Algorithms for Materials

Design: Evolutionary Algorithms That Learn. ACS Comb. Sci. 2017,

19,96−107.

(37) Algara-Siller, G.; Kurasch, S.; Sedighi, M.; Lehtinen, O.; Kaiser,

U. The Pristine Atomic Structure of MoS2Monolayer Protected from

Electron Radiation Damage by Graphene. Appl. Phys. Lett. 2013, 103,

203107.

(38) Cheng, P.; Sun, K.; Hu, Y. H. Mechanically-Induced Reverse

Phase Transformation of MoS2 from Stable 2H to Metastable 1T and

Its Memristive Behavior. RSC Adv. 2016, 6, 65691−65697.

(39) Eda, G.; Yamaguchi, H.; Voiry, D.; Fujita, T.; Chen, M.;

Chhowalla, M. Photoluminescence from Chemically Exfoliated MoS2.

Nano Lett. 2011, 11, 5111−5116.

(40) Hong, J.; Pan, Y.; Hu, Z.; Lv, D.; Jin, C.; Ji, W.; Yuan, J.; Zhang,

Z. Direct Imaging of Kinetic Pathways of Atomic Diffusion in

Monolayer Molybdenum Disulfide. Nano Lett. 2017, 17, 3383−3390.

(41) Komsa, H.-P.; Kurasch, S.; Lehtinen, O.; Kaiser, U.;

Krasheninnikov, A. V. From Point to Extended Defects in Two-

Dimensional MoS2: Evolution of Atomic Structure under Electron

Irradiation. Phys. Rev. B: Condens. Matter Mater. Phys. 2013, 88,

035301.

(42) Saha, D.; Mahapatra, S. Atomistic Modeling of the Metallic-to-

Semiconducting Phase Boundaries in Monolayer MoS2. Appl. Phys.

Lett. 2016, 108, 253106.

(43) Masubuchi, S.; Morimoto, M.; Morikawa, S.; Onodera, M.;

Asakawa, Y.; Watanabe, K.; Taniguchi, T.; Machida, T. Autonomous

Robotic Searching and Assembly of Two-Dimensional Crystals to

Build van Der Waals Superlattices. Nat. Commun.

2018, 9, 1413.

(44) Lin, X.; Si, Z.; Fu, W.; Yang, J.; Guo, S.; Cao, Y.; Zhang, J.;

Wang, X.; Liu, P.; Jiang, K.; Zhang, Y.; Zhao, W. Intelligent

Identiﬁcation of Two-Dimensional Structure by Machine-Learning

Optical Microscopy. ArXiv180302062 Cond-Mat 2018.

(45) Wang, C. C.; Pilania, G.; Boggs, S. A.; Kumar, S.; Breneman,

C.; Ramprasad, R. Computational Strategies for Polymer Dielectrics

Design. Polymer 2014, 55, 979−988.

(46) Zhu, Q.; Sharma, V.; Oganov, A. R.; Ramprasad, R. Predicting

Polymeric Crystal Structures by Evolutionary Algorithms. J. Chem.

Phys. 2014, 141, 154102.

(47) Meenakshisundaram, V.; Hung, J.-H.; Patra, T. K.; Simmons,

D. S. Designing Sequence-Specific Copolymer Compatibilizers Using

a Molecular-Dynamics-Simulation-Based Genetic Algorithm. Macro-

molecules 2017, 50, 1155 −1166.

(48) Botu, V.; Batra, R.; Chapman, J.; Ramprasad, R. Machine

Learning Force Fields: Construction, Validation, and Outlook. J. Phys.

Chem. C 2017, 121, 511−522.

(49) Cherukara, M. J.; Narayanan, B.; Kinaci, A.; Sasikumar, K.;

Gray, S. K.; Chan, M. K. Y.; Sankaranarayanan, S. K. R. S. Ab Initio-

Based Bond Order Potential to Investigate Low Thermal Con-

ductivity of Stanene Nanostructures. J. Phys. Chem. Lett. 2016, 7,

3752−3759.

(50) Narayanan, B.; Chan, H.; Kinaci, A.; Sen, F. G.; Gray, S. K.;

Chan, M. K. Y.; Sankaranarayanan, S. K. R. S. Machine Learnt Bond

Order Potential to Model Metal−Organic (Co−C) Heterostructures.

Nanoscale 2017, 9, 18229−18239.

(51) Zhang, Z.; Schwanz, D.; Narayanan, B.; Kotiuga, M.; Dura, J.

A.; Cherukara, M.; Zhou, H.; Freeland, J. W.; Li, J.; Sutarto, R.; He,

F.; Wu, C.; Zhu, J.; Sun, Y.; Ramadoss, K.; Nonnenmann, S. S.; Yu,

N.; Comin, R.; Rabe, K. M.; Sankaranarayanan, S. K. R. S.; et al.

Perovskite Nickelates as Electric-Field Sensors in Salt Water. Nature

2018, 553,68−72.

(52) Erdemir, A.; Ramirez, G.; Eryilmaz, O. L.; Narayanan, B.; Liao,

Y.; Kamath, G.; Sankaranarayanan, S. K. R. S. C arbon-Based

Tribofilms from Lubricating Oils. Nature 2016, 536,67−71.

(53) Asadi, M.; Sayahpour, B.; Abbasi, P.; Ngo, A. T.; Karis, K.;

Jokisaari, J. R.; Liu, C.; Narayanan, B.; Gerard, M.; Yasaei, P.; Hu, X.;

Mukherjee, A.; Lau, K. C.; Assary, R. S.; Khalili-Araghi, F.; Klie, R. F.;

Curtiss, L. A.; Salehi-Khojin, A. A Lithium

−Oxygen Battery with a

Long Cycle Life in an Air-like Atmosphere. Nature 2018, 555, 502−

506.

(54) Bahns, J. T.; Sankaranarayanan, S. K. R. S.; Gray, S. K.; Chen,

L. Optically Directed Assembly of Continuous Mesoscale Filaments.

Phys. Rev. Lett. 2011, 106, 095501.

(55) Bahns, J. T.; Sankaranarayanan, S. K. R. S.; Giebink, N. C.;

Xiong, H.; Gray, S. K. Optically Directed Mesoscale Assembly and

Patterning of Electrically Conductive Organic−Inorganic Hybrid

Structures. Adv. Mater. 2012, 24, OP242−OP246.

(56) Deringer, V. L.; Bernstein, N.; Barto

k, A. P.; Cliffe, M. J.;

Kerber, R. N.; Marbella, L. E.; Grey, C. P.; Elliott, S. R.; Csa

nyi, G.

Realistic Atomistic Structure of Amorphous Silicon from Machine-

Learning-Driven Molecular Dynamics. J. Phys. Chem. Lett. 2018, 9,

2879.

(57) Xiang, H. J.; Wei, S.-H.; Gong, X. G. Structural Motifs in

Oxidized Graphene: A Genetic Algorithm Study Based on Density

Functional Theory. Phys. Rev. B: Condens. Matter Mater. Phys. 2010,

82, 035416.

(58) Gu, T.; Luo, W.; Xiang, H. Prediction of Two-Dimensional

Materials by the Global Optimization Approach: Prediction of Two-

Dimensional Materials by the Global Optimization Approach. WIREs

Comput. Mol. Sci. 2017, 7, No. e1295.

(59) Sanz-Navarro, C. F.; Åstrand, P.-O.; Chen, D.; Rønning, M.;

van Duin, A. C. T.; Goddard, W. A. Molecular Dynamics Simulations

of Metal Clusters Supported on Fishbone Carbon Nanofibers. J. Phys.

Chem. C 2010, 114 , 3522−3530.

(60) Hashimoto, A.; Suenaga, K.; Gloter, A.; Urita, K.; Iijima, S.

Direct Evidence for Atomic Defects in Graphene Layers. Nature 2004,

430, 870−873.

(61) Meyer, J. C.; Girit, C. O.; Crommie, M. F.; Zettl, A. Imaging

and Dynamics of Light Atoms and Molecules on Graphene. Nature

2008, 454, 319−322.

(62) Kinaci, A.; Narayanan, B.; Sen, F. G.; Davis, M. J.; Gray, S. K.;

Sankaranarayanan, S. K. R. S.; Chan, M. K. Y. Unraveling the Planar-

Globular Transition in Gold Nanoclusters through Evolutionary

Search. Sci. Rep. 2016, 6, 34974.

(63) Chua, A. L.-S.; Benedek, N. A.; Chen, L.; Finnis, M. W.; Sutton,

A. P. A Genetic Algorithm for Predicting the Structures of Interfaces

in Multicomponent Systems. Nat. Mater. 2010, 9, 418−422.

(64) Plimpton, S. Fast Parallel Alg orithms for Short-Range

Molecular Dynamics. J. Comput. Phys. 1995, 117,1−19.

(65) Martyna, G. J.; Klein, M. L.; Tuckerman, M. Nose

−Hoover

Chains: The Canonical Ensemble via Continuous Dynamics. J. Chem.

Phys. 1992, 97, 2635−2643.

(66) Das, S.; Bera, M. K.; Tong, S.; Narayanan, B.; Kamath, G.;

Mane, A.; Paulikas, A. P.; Antonio, M. R.; Sankaranarayanan, S. K. R.

S.; Roelofs, A. K. A Self-Limiting Electro-Ablation Technique for the

Top-Down Synthesis of Large-Area Monolayer Flakes of 2D

Materials. Sci. Rep. 2016, 6, 28195.

(67) Huang, Y.-T.; Dodda, A.; Schulman, D. S.; Sebastian, A.;

Zhang, F.; Buzzell, D.; Terrones, M.; Feng, S.-P.; Das, S. Anomalous

Corrosion of Bulk Transition Metal Diselenides Leading to Stable

Monolayers. ACS Appl. Mater. Interfaces 2017, 9, 39059−39068.

(68) Schulman, D. S.; Sebastian, A.; Buzzell, D.; Huang, Y.-T.;

Arnold, A. J.; Das, S. Facile Electrochemical Synthesis of 2D

ACS Nano Article

DOI: 10.1021/acsnano.8b02844

ACS Nano 2018, 12, 8006−8016

8015

Monolayers for High-Performance Thin-Film Transistors. ACS Appl.

Mater. Interfaces 2017, 9, 44617−44624.

(69) Schulman, D. S.; May-Rawding, D.; Zhang, F.; Buzzell, D.;

Alem, N.; Das, S. Superior Electro-Oxidation and Corrosion

Resistance of Monolayer Transition Metal Disulfides. ACS Appl.

Mater. Interfaces 2018, 10, 4285−4294.

(70) Zhang, F.; Erb, C.; Runkle, L.; Zhang, X.; Alem, N. Etchant-

Free Transfer of 2D Nanostructures. Nanotechnology 2018, 29,

025602.

ACS Nano Article

DOI: 10.1021/acsnano.8b02844

ACS Nano 2018, 12, 8006−8016

8016